docs: correct croissant.json

croissant.json

@@ -100,7 +100,7 @@
   "citeAs": "@inproceedings{liu2026openclawbench, title={{OpenClawBench}: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories}, author={Liu, Yibing and Liu, Yangze and Yin, Xiaolong and Wang, Bin and Zhang, Chong and Yin, Hao and Han, Zhongyi}, booktitle={Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track}, year={2026}}",
   "isLiveDataset": false,
   "rai:dataLimitations": "Silver labels come from a single LLM judge and contain errors; evaluation claims should use the human_verified subset. Subtype boundaries (notably weak_evidence_commitment vs. capability_gap_overcommitment) are inherently fuzzy and are documented in TAXONOMY.md. Only six open-weight backbones on BFCL function-calling tasks are covered; results may not transfer to proprietary agents or open-ended, long-horizon tasks. The supp140 human subset is stratified by silver subtype, so its anomaly rate is not representative of the corpus.",
-  "rai:dataBiases": "Source-model bias: six open-weight backbones (DeepSeek-R1, GLM-4.5-Air, GPT-OSS-20B, Qwen3.5-9B, Qwen3.6-27B, Qwen3.6-35B-A3B), four of them from one model family; no proprietary models. Task bias: BFCL function-calling prompts, which emphasize structured API invocation. Language bias: English only. Label bias: silver labels reflect the judge model's and the annotation prompt's notion of commitment under information deficit.",
+  "rai:dataBiases": "Source-model bias: six open-weight backbones (DeepSeek-R1, GLM-4.5-Air, GPT-OSS-20B, Qwen3.5-9B, Qwen3.6-27B, Qwen3.6-35B-A3B), three of them (Qwen) from one model family; no proprietary models. Task bias: BFCL function-calling prompts, which emphasize structured API invocation. Language bias: English only. Label bias: silver labels reflect the judge model's and the annotation prompt's notion of commitment under information deficit.",
   "rai:personalSensitiveInformation": "None. Trajectories are model outputs on public BFCL prompts and contain no personal data. API keys and access tokens appearing in trajectory text (the agents' own runtime credential and third-party key-like strings from fetched web pages or agent-written code) were replaced with placeholders.",
   "rai:dataUseCases": "Training and evaluating trajectory-level anomaly detectors; subtype classification; onset localization; studying LLM-judge reliability against human labels; comparing agent backbones on process-side reliability. Not intended for drawing conclusions about proprietary agents or for automated monitoring of end users without human review.",
   "rai:dataSocialImpact": "Supports more reliable deployment of tool-using LLM agents by making process failures (acting or answering confidently without adequate information) measurable and trainable. Misuse risk is low: the data contain no personal information and describe failure patterns rather than harmful capabilities.",